World Models in Pieces: Structural Certification for General Agents

The paper introduces a novel approach called structural certification for general agents, addressing the limitations of standard worst-case analysis in the big-world regime. By formalizing the concept that general agents cannot be universally capable, the authors present algorithms that use deep compositional goals to filter transitions, achieving an error bound of $\mathcal{O}(1/n) + \mathcal{O}(\delta)$ for goal-conditioned performance. This framework allows practitioners to certify the reliability of long-horizon planning in specific transitions, enhancing the deployment of general agents in complex environments.